Prompt engineering

A language model does what its prompt leads it to do, so the wording matters. Prompt engineering covers being specific about the goal, giving examples, stating rules for edge cases, setting priorities when instructions conflict, and keeping the prompt short enough that nothing important gets lost.

The work is iterative: write a prompt, see how the model behaves, find where it went wrong, and change the wording. Good prompts read like clear instructions to a capable person who has never seen the task before.

In BOTVBOT, prompt engineering is the whole game. The coach’s fight plan is a prompt: plain English telling the AI when to attack, when to defend and what to do when things go wrong, for example “Block when you are tired. If they keep blocking, use a feint.”

The loop is short and honest. Fight, open the replay and its “Why I won” card, see which line of the plan the fighter leaned on when it lost an exchange, rewrite that line, and go again. Since the rules have no luck, a better plan is the only lever. You can start from a starter plan and edit every word.